An Efficient Method for Big Data Classification using Optimized Error-Minimized and Ridge-Regression ELM

A parallelized learning algorithm called Extreme Learning Machine tree (ELM) is used to classify the data instances more effectively in big data analysis. In ELM Tree approach, the over-partitioning problem in Decision Tree (DT) generation is resolved by using ELM as a leaf node when the information gain ratio or ambiguity of available cuts are smaller than a specified threshold. However, perturbation, multicollinearity, and stability are other shortcomings that constrain the performance of ELM. This research proposes an Optimized Error-minimized and Ridge-regression Extreme Learning Machine (OER-ELM) to address those challenges. In OER-ELM, the required number of hidden nodes during learning stage is incremented to obtain minimum error. The least-square method-based output weight calculation is replaced by ridge-regression. The initially chosen input weights and hidden biases of ELM are typically selected at random. This work proposes techniques for optimization such as Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Firefly (FF) algorithm to modify the input weights and hidden biases. The experiment is conducted for GA- ELM Tree, GA-OER-ELM Tree, PSO-ELM Tree, PSO-OER-ELM Tree, FF-ELM Tree and FF-OER-ELM Tree on three big datasets. Based on the results of the experiments, it was determined that OER-ELM Tree is superior to ELM Tree in terms of classification accuracy and computing efficiency.